> We got the following running times...
> 1 node - 13hrs 5:27
> 2 nodes - 16hrs 12:59
> 4 nodes - 17hrs 32:14
> 8 nodes - 45hrs 00:36
> 16 nodes - 45hrs 46:55
Hi,
I don't know anything about GROMACS but a reasons for the behavior as you
described of parallel programs could be the following.
the more nodes the smaller the pice of calculation for each node and the more
communication between the processes. So, if the problem is small calculation
of the whole problem by one node without communication is much faster then
calculation of smaller pieces by severel nodes which have to communicate with
each other. (communication is mostly much more time expensive than
calculation)
Try to increase the calculation complexity (e.g. a biger input matrix). If the
problem is the one I described this should make the times a bit more equal to
each other.
In the issue I think only studying the source code would give a detailed
answer.
hope that helps a bit.
--
Tobias Wenzel
Technische Universität Chemnitz (germany)
|
